Right, and I mentioned CUDA earlier as one of the reason of their success, so it’s definitely something important. Clients might be interested in e.g Google TPU, startups like Etched, Tenstorrent, Groq, Cerebras Systems or heck even design their own but are probably limited by their current stack relying on CUDA. I imagine though that if backlog do keep on existing there will be abstraction libraries, at least for the most popular ones e.g TensorFlow, JAX or PyTorch, simply because the cost of waiting is too high.
Anyway what I meant isn’t about hardware or software but rather ROI, namely when Goldman Sachs and others issue analyst report saying that the promise itself isn’t up to par with actual usage for paying customers.
Why do you believe that? As far as I understand, other HW exists…but no SW to run on it…
Right, and I mentioned CUDA earlier as one of the reason of their success, so it’s definitely something important. Clients might be interested in e.g Google TPU, startups like Etched, Tenstorrent, Groq, Cerebras Systems or heck even design their own but are probably limited by their current stack relying on CUDA. I imagine though that if backlog do keep on existing there will be abstraction libraries, at least for the most popular ones e.g TensorFlow, JAX or PyTorch, simply because the cost of waiting is too high.
Anyway what I meant isn’t about hardware or software but rather ROI, namely when Goldman Sachs and others issue analyst report saying that the promise itself isn’t up to par with actual usage for paying customers.